Principia: Test-Time Hypotheses to Principle Construction for Scientific Discovery

Principia. Test-Time Hypotheses to Principle Construction for Scientific Discovery. By Dongki Kim and four others, from KAIST and DeepAuto AI. Published on the twenty-third of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Scientific discovery depends on identifying the principles that underlie observed phenomena, because those principles guide what to investigate next. In complex discovery processes, the empirical principles are often unknown. They have to be constructed through a hypotheses-to-principle pipeline, which extracts recurring causalities from observations that accumulate progressively during test-time discovery. That construction demands substantial domain expertise, time and effort, and it makes this translation stage a bottleneck for discovery as a whole. Observations can pile up faster than anyone can turn them into general statements about what causes what, and until that happens, the observations cannot steer the next round of work.

Principia is an LLM-based framework that constructs scientific principles at test time, and keeps refining them. It draws on how human researchers develop principles. From related inputs and their observed outcomes, it forms comparative causal hypotheses. It then synthesizes the relationships that recur across those hypotheses into inductive principles. Hypotheses are the specific layer; principles are what survives across them. The order mirrors the bottleneck it is meant to relieve. First explain particular outcomes by comparing them, then generalise from the explanations that keep coming back.

A loop keeps that knowledge moving. Principia applies the hypotheses and principles it has built to the task, and obtains new observations. It revises the hypotheses it used in light of those observations. It then reconstructs the principles from the updated hypotheses. Each pass feeds the next, so the scientific knowledge continues to evolve throughout test-time discovery, instead of being settled once at the start.

The benchmarks span molecular optimization, protein engineering, gene perturbation and molecular property prediction. Across them, Principia outperforms three kinds of baseline: those relying on internal knowledge, those drawing on external sources, and those built on experience summaries. Beyond the scores, the principles it generates are plausible and consistent with the scientific literature. The abstract gives no figures for either result. The two claims answer different worries: one that the principles help the task, the other that they are sound science.

What Principia takes over is the step between seeing results and stating why they happen. Hypotheses come from comparisons between related cases. Principles come from what those hypotheses share. And both are rebuilt as new observations arrive, so the account of the task stays tied to the latest evidence rather than to whatever was believed when the search began.